winch(aarch64): Revisit the shadow stack pointer approach

[saulecabrera]

This commit marks another step toward finalizing AArch64 support in
Winch.

While enabling spec tests, I experienced some unexpected failures
related to Wasm loads/stores and traps. The observed
symptoms are as follows:

• Under normal conditions, Wasm loads/stores work as expected.
• In out-of-bounds scenarios, loads/stores result in a segmentation
  fault, whereas the expected behavior is to trigger an out-of-bounds trap.
• When out-of-bounds access can be determined statically, the program
  still results in a segmentation fault instead of the anticipated
  out-of-bounds trap.

Debugging revealed the following issues:

• The stack pointer was not correctly aligned to 16 bytes when entering
  signal handlers, which caused the segmentation fault.
• Wasm loads and stores were not flagged as untrusted, leading to
  segmentation faults even when the stack pointer was properly aligned.

This commit fixes the previous issues by:

• Correctly flagging wasm loads and stores as untrusted.
• Reworking the shadow stack pointer approach such that it allows
  aligning the stack pointer at arbitrary points in the program,
  particularly where signal handling might be needed. This rework
  involves changing some principles introduced in
  winch: Use aarch64 backend for code emission. #5652; namely:
  changing the primary stack pointer register to be the shadow stack
  pointer. See the updated comments in the code for more details.

Note that this change doesn't enable spectests. I'll follow-up with more work to do so. To try this change, run:

  cargo run -- wast -Ccompiler=winch tests/spec_testsuite/address.wast

--

The diff is large-ish due to all the changes in disassembly tests, however, all code changes can be found in this commit.

[alexcrichton]

While I'm not expert on Winch this all looks and sounds reasonable enough to me 👍